---

name: boxmot description: "Use BoxMOT for multi-object tracking, benchmark replay, ReID model lifecycle workflows, and native C++ tracker backends." metadata: disco-role: operating disable-model-invocation: true license: AGPL 3.0

BoxMOT

BoxMOT is the skill to use when the user wants to track objects, run MOT benchmarks, train or compare ReID models, export ReID checkpoints, or switch trackers to native C++ backends.

Quick install and smoke check

For a public install, use the package itself:

pip install boxmot
boxmot --help
python -c "import boxmot; print(boxmot.__version__)"

If you are working from a source checkout, uv sync --all-extras --all-groups is the broadest dev-friendly install, but the runtime skill should stay usable from the published package too.

If the CLI or import fails, read:

• install and inspect
• cross-cutting troubleshooting
• repo provenance

Route map

tracking-and-trackers

Use this route for direct tracking, tracker construction, tracker defaults, AABB/OBB detection layouts, output schemas, and tracker-level debugging.

Typical prompts:

• track a webcam or video with a specific tracker
• explain tracker output columns or OBB input shapes
• instantiate a tracker from Python and inspect its results
• debug shape errors or missing det_ind values

Read:

• sub-skills/tracking-and-trackers/SKILL.md
• sub-skills/tracking-and-trackers/references/workflows.md
• sub-skills/tracking-and-trackers/references/data-formats.md

benchmark-workflows

Use this route for cached generate / eval / tune / research workflows, benchmark YAMLs, cache reuse, public detections, and postprocessing.

Typical prompts:

• generate caches for a benchmark
• reuse an existing cache with a different tracker
• tune tracker hyperparameters on MOT17 or MMOT
• debug benchmark config selection or split handling

Read:

• sub-skills/benchmark-workflows/SKILL.md
• sub-skills/benchmark-workflows/references/benchmark-workflows.md
• sub-skills/benchmark-workflows/references/configuration.md

reid-lifecycle

Use this route for training ReID backbones, evaluating checkpoints, comparing multiple checkpoints, exporting to ONNX/OpenVINO/TensorRT/TFLite, and generating embeddings.

Typical prompts:

• train a ReID model on a dataset or recipe
• evaluate or compare checkpoints across target datasets
• export a ReID model for deployment
• debug dataset, preprocess, or export-format issues

Read:

• sub-skills/reid-lifecycle/SKILL.md
• sub-skills/reid-lifecycle/references/reid-lifecycle.md
• sub-skills/reid-lifecycle/references/model-and-export-overview.md

native-cpp-backends

Use this route for --tracker-backend cpp, boxmot build, native C++ live tracking, native replay, and embedding the trackers in your own C++ program.

Typical prompts:

• build the native trackers
• switch a supported tracker to the C++ backend
• understand supported native tracker coverage
• troubleshoot build-tool or backend-selection failures

Read:

• sub-skills/native-cpp-backends/SKILL.md
• sub-skills/native-cpp-backends/references/native-cpp.md
• sub-skills/native-cpp-backends/references/troubleshooting.md

Public package facts

• CLI entry point: boxmot
• Public Python facade: boxmot.BoxMOT
• Public detector wrapper: boxmot.Detector
• Public ReID wrapper: boxmot.ReIDModel
• Package-level tracker export: boxmot.trackers.OccluBoost
• Supported tracker names are defined in boxmot.trackers.registry.TRACKER_MAPPING
• Detection layouts switch automatically from tensor shape:
  • AABB: (N, 6) input -> (N, 8) output
  • OBB: (N, 7) input -> (N, 9) output

When to start here

If the user only says "use BoxMOT", start with the route that matches the task family and then open the nearest sub-skill reference. If the request names a CLI flag, benchmark id, tracker name, ReID checkpoint, output schema, or C++ backend, use that signal to choose the route immediately.